--- title: "What $1.4 trillion is supposed to buy" publication: deslop.media byline: deslop.media editorial provenance: editorial_synthesis provenance_note: "Editorial synthesis by deslop.media editorial, assembled from the subject's public record. The subject did not sit for an interview with deslop.media and has not reviewed it." issue: 1 issue_title: The Scaling Bet grade: draft issue_version: 3 issue_version_date: 2026-10-11 piece_version: 1 first_published: 2026-10-11 reporting_current_through: 2026-10 canonical_url: https://deslop.media/feature/sam-altman/the-stargate-bet/ cite_as: "\"What $1.4 trillion is supposed to buy,\" deslop.media editorial, Issue 1 (Draft v1, October 11, 2026)" changes: https://deslop.media/issues/1-the-scaling-bet/changes/ flag: "MCP tool flag_content, or POST https://api.deslop.media/api/flags" --- On January 21, 2025, the day after his second inauguration, President Trump stood in the Roosevelt Room flanked by Sam Altman, Masayoshi Son of SoftBank, and Larry Ellison of Oracle. They announced "The Stargate Project," a joint venture pledging $500 billion over four years to build AI infrastructure inside the United States. SoftBank and OpenAI took roughly equal stakes; Oracle was named primary technology partner; MGX, the Abu Dhabi sovereign-wealth vehicle, joined as the fourth principal. An initial $100 billion was to deploy immediately, beginning with a campus already under construction in Abilene, Texas. By Ellison's telling to Reuters that week, Abilene would grow to twenty data halls of roughly half a million square feet each. The architects of the science were not in the room. Jared Kaplan, co-author of the 2020 scaling-laws paper, and Jordan Hoffmann, who led the 2022 Chinchilla paper that refined it, were both absent from the announcement of the $500 billion pledge. Son, Ellison, and the sovereign-wealth principal from Abu Dhabi were the actual deal. The curve was assumed; what was being assembled, in front of cameras, was the body of capital capable of carrying it. ## The bet beneath the bet Underneath the announcement is a wager Altman has been making, in plain sight, since 2024. It runs in four steps. If OpenAI can get the compute, then capability scales with that compute; then the cost of a given level of intelligence falls with time and with the surrounding ecosystem, and some of that benefit accrues to the lab that built it; then demand grows faster than the falling price, because cheaper, better intelligence gets used far more. The conclusion Altman draws from the four steps is that the compute is worth buying — and that a company that buys the most of it, fastest, becomes a good business. Everything starts with the first step. The other three are claims about how the world behaves once the chips are in the racks; the first is a claim Altman has to make true by going out and doing it. The rest of this account takes the three claims about the world first, then the chase for the compute that the whole structure rests on. ## Performance scales with compute The first claim is that capability rises predictably as compute rises, and on the public record it has held. Altman set the proposition down himself. In *The Intelligence Age*, in September 2024, he stated the empirical fact plainly: "deep learning worked, got predictably better with scale, and we dedicated increasing resources to it." In *Three Observations*, in February 2025, he stated it as something close to a law — the intelligence of an AI model "roughly equals the log of the resources used to train and run it." The frontier results since have behaved accordingly. Gemini 3 Deep Think, the Q1 2026 update to Google DeepMind's reasoning system, scored 93.8% on GPQA Diamond, a graduate-level science benchmark, and reached an Elo of 3455 on Codeforces, in the top tier of competitive programming. An earlier model had already set the mathematics mark. On July 21, 2025, Google DeepMind announced that an advanced version of Gemini Deep Think, built on Gemini 2.5, had earned a gold-medal score of 35 out of 42 at the International Mathematical Olympiad, certified by IMO President Gregor Dolinar. It solved five of six problems end-to-end in natural language inside the 4.5-hour competition window. The sterner test is the benchmarks built after the curve started embarrassing the old ones. ARC-AGI-2, the reasoning evaluation François Chollet and his collaborators released in March 2025 specifically to resist memorization, launched with pure language models scoring zero and the best public reasoning systems in single digits. By February 2026, Google's Gemini 3 Deep Think had reached 84.6% on it, verified by ARC Prize on the Semi-Private set. METR's Time Horizon 1.1, published January 29, 2026 by the evaluation outfit Model Evaluation and Threat Research, reports a doubling time of 130.8 days, for models released since 2023, in its 50% time horizon: the length of task, timed by how long skilled humans take, that a model completes with 50% success. That is faster than the seven-month long-run trend the same group had previously tracked. The capability story is not frictionless. The strongest published critique of the inference-time-compute version of it, "Inverse Scaling in Test-Time Compute" (arXiv 2507.14417, July 2025), finds the curve real but identifies five failure modes where more thinking-time degrades rather than improves an answer — distractor-attention drift, framing-overfit, spurious-correlation drift, focus loss on long deductive chains, and amplified self-preservation under extended thinking. The paper sharpens the claim rather than overturning it. The science is, broadly, behaving the way Altman said it would. ## Cost declines with time and the ecosystem The second claim is that the cost of a given level of intelligence falls fast — and that some of the benefit accrues to the lab that built the frontier. It has held in the way Altman framed it, and bent in a way he did not. His own version, again from *Three Observations*, is that the cost of using a given level of AI "falls about 10x every 12 months, and lower prices lead to much more use." The third-party data puts the rate in the same range. Hans Gundlach and colleagues, in "The Price of Progress" (arXiv 2511.23455, revised March 2026), found that the price for a given level of benchmark performance has decreased "remarkably fast, around 5× to 10× per year, for frontier models on knowledge, reasoning, math, and software engineering benchmarks." On GPQA Diamond, the top-bin price decline runs roughly 31x per year, the strongest single datapoint behind Altman's claim. The part of the ecosystem that complicates the claim is the open-weight competition. DeepSeek, the Chinese lab whose R1 model launched on January 20, 2025 — the day before the Roosevelt Room — at frontier-comparable benchmark scores and a small fraction of the leading per-token price, has continued to reset the price floor. Its V4-Flash, released in April 2026, ran at $0.14 per million uncached input tokens and $0.28 per million output tokens, roughly 36x cheaper on input than GPT-5.5's standard rate of $5 per million. The ratio compares list prices per token, not capability or cost per task. DeepSeek retired it on September 10, 2026, and moved its traffic to V4.1-Flash under new pricing. The benefit of the falling cost curve, on this evidence, does not accrue only to the lab that built the frontier; an open-weight follower can price most of it away within months. Altman's claim that cost falls is amply confirmed. Whether the lab that builds the frontier keeps the value of the fall is the question DeepSeek presses. Gundlach and colleagues add a second complication that *Three Observations* did not contemplate. While per-token costs fall at the projected rate, per-task costs at the frontier are rising, because reasoning models burn tokens to think and agentic loops run them in long chains. Menlo Ventures, a venture-capital firm, estimated in December 2025 that enterprise spending on generative AI rose from $11.5 billion in 2024 to $37 billion in 2025, even as per-token prices fell. Its count covers models, training infrastructure and applications, and excludes inference and model serving on clouds such as AWS, Google Cloud and Azure. The per-token curve and the per-task frontier curve point in opposite directions on the same axis. The falling cost is real; it also keeps incentivizing the capacity build that cheaper tokens were supposed to make unnecessary. ## Demand grows faster than the price falls The third claim is that demand is elastic enough that it outruns the falling price — as performance rises and cost drops, intelligence gets used far more than proportionally. Altman stated it at its strongest in *Three Observations*: the socioeconomic value of linearly increasing intelligence is "super-exponential in nature," and "exponentially increasing investment" should therefore not stop in the near term. The proposition is not that demand will absorb whatever the buildout produces. It is that supply and demand compound together, so that capital spent against the curve compounds with them. The empirical record is mixed. ChatGPT's weekly active users moved from 100 million in late 2023 to 300 million by January 2025, to 800 million at OpenAI's DevDay in October 2025, to 900 million by February 27, 2026 — a ninefold rise in a little over two years that, by any consumer-internet measure, is an exceptional adoption curve. By April 2026, data from Sensor Tower, an app-analytics firm, showed ChatGPT downloads growing only about 14% year-over-year while Claude downloads grew roughly tenfold from a smaller base; ChatGPT uninstalls rose 132% year-over-year in the same window. On August 31, 2026, OpenAI said ChatGPT had passed one billion weekly active users, and that ChatGPT Ads, launched less than 200 days earlier, had reached a $1 billion annualized revenue run rate by the company's count. Altman's answer is that the consumer app is the wrong denominator, and that the developer interface is where the elasticity argument is actually being run. OpenAI claimed four million developers and six billion tokens per minute at DevDay 2025. Spending on the models themselves falls inside the enterprise generative-AI market that Menlo Ventures estimated grew 3.2-fold in 2025, to $37 billion. Anthropic, the rival lab building on the same scaling premise, is growing on that axis too: its run-rate revenue passed $30 billion in April 2026, per The Register, and on May 6, 2026 it reported a roughly 80-fold annualized growth rate in the first quarter of 2026 (Fortune, May 8, 2026). The throughput is real and growing the way the elasticity argument predicts; the slowdown in ChatGPT downloads through April 2026 was also real. Both are evidence about the same market, on different time horizons. ## The chase for compute The three claims describe how the world is supposed to behave once the compute exists. Getting it to exist is the part Altman cannot assume, and it is the part he has been chasing since January 2025. OpenAI's president, Greg Brockman, put the company's posture plainly on the Big Technology Podcast in April 2026. Recalling a conversation from the launch of ChatGPT, he said he was asked how much compute the company should buy, and answered: "All of it." The line is the institutional ambition Altman is executing against — the working assumption that demand for intelligence is effectively unbounded, so the correct acquisition posture is maximal. The quest that follows is the attempt to act on it. Microsoft came first. The 2019 partnership that gave OpenAI its early compute had hardened, by 2024, into two contractual anchors: a cloud-exclusivity arrangement that routed OpenAI's capacity through Azure, and a clause that would terminate Microsoft's access rights if OpenAI declared it had reached AGI. Through 2025 and into 2026 both were amended out, and the relationship was reorganized around Microsoft as an investor in OpenAI's new Public Benefit Corporation rather than as its exclusive channel. The reframing relieved a real constraint. With exclusivity gone, OpenAI could meet enterprise customers on the clouds they already used; the company told employees, in documents surfaced during the Musk litigation, that the old arrangement "limited our ability to meet enterprises where they are." The April 2026 cap renegotiation that retired the AGI clause was the culmination — the contractual signature of the founding relationship traded for room to maneuver, with Microsoft compensated as a shareholder rather than a gatekeeper. From Microsoft the chase widened into a web. Each new counterparty added compute, and the deals share a signature: the supplier's economics are coupled to OpenAI's deployment volume rather than settled up front. AMD granted warrants for up to 160 million shares — a roughly ten percent economic stake at a penny apiece — that vest as OpenAI buys chips. NVIDIA's September 2025 letter of intent keyed a progressive investment of up to $100 billion to the gigawatts OpenAI deploys; in the funding round OpenAI announced on February 27, 2026, NVIDIA committed $30 billion of equity, alongside dedicated OpenAI capacity on NVIDIA's Vera Rubin systems. Oracle holds a multi-year contract reported at $300 billion; AWS, Broadcom, CoreWeave, and a $1 billion working-capital line OpenAI extended to Cerebras fill in the rest. The structure turns each supplier into a stakeholder, which is how a company without hyperscaler cash flow gets hyperscaler-scale capacity committed to it. The chase has a harder edge, and it is a business one. Altman's signature as a deal-maker is that he renegotiates terms after they are struck, as conditions shift. Six commitments inside the announced pile were walked back, restructured, or relocated between October 2025 and April 2026, and the pattern is consistent: OpenAI declines a tranche, and Microsoft, the partner whose balance sheet can most readily hold it, absorbs the capacity. Norway is the plainest instance. On April 15, 2026, OpenAI dropped its planned 230-megawatt offtake from the Narvik facility operated by Nscale, a UK-based GPU-cloud operator; Microsoft took over the capacity and contracted more than 30,000 NVIDIA Rubin GPUs at the same campus. Read one at a time, each move has its own rationale — energy economics in the UK, a contractual disagreement at Narvik, expansion-cost discipline at Abilene. Read together, they are the same maneuver the AGI-clause renegotiation was: agreed terms reopened when the conditions behind them changed. The same Microsoft characterized in 2024 as the cautious party to the relationship ends the window holding more gigawatts than it did at the start. Re-trading is how the program is run. A program mobilizing capital at this scale against a four-year clock is continuously repriced as chips, power, and demand move underneath it, and Altman treats every signed commitment as renegotiable when the numbers move. The ambition that opened the chase is also its constraint. As Brockman put it in the same April 2026 conversation, the demand is such that "there just isn't enough compute in the world" to satisfy it — which is why the search never settles, and why the terms keep moving. ## The focus turn The chase for compute has a mirror image inside the company: a sharpening of where that compute, and the revenue meant to justify it, are pointed. Through the first half of 2026, OpenAI narrowed visibly toward its enterprise and developer business and away from the consumer experiments around it. The pivot is on the record. CFO Sarah Friar said in March 2026 that enterprise clients account for "over 40%" of OpenAI's revenue, and that she expects the figure to reach 50% by year-end. The revenue picture was less firm: in May 2026, analyst estimates for OpenAI's 2026 revenue spanned a wide range, from roughly $25 billion to the $34 billion that Reuters Breakingviews, citing HSBC, reported — projections, not audited figures. Fidji Simo, who runs OpenAI's applications business, framed the narrowing to staff in terms a competitor would recognize, reportedly telling them, per the Wall Street Journal and The Verge: "We cannot miss this moment because we are distracted by side quests." The clearest side quest to end was Sora. In March 2026 OpenAI announced that it would discontinue its standalone AI video app and the Sora API, an OpenAI spokesperson, Kayla Wood, confirming the shutdown on the record; the app closed on April 26 and the API on September 24. The move also wound down a $1 billion partnership with Disney. Against the things it cut, OpenAI pushed its coding agent, Codex, first released in May 2025 and given a desktop app in February 2026, to close the lead Anthropic's Claude Code had opened in software development. The early traction is company-stated: OpenAI says most of its technical staff now use Codex. The narrowing is where the demand claim meets the chase for compute: the company choosing where the compute, and the revenue, actually go. ## The Altman bet As a number, OpenAI's announced compute commitment is roughly $1.4 trillion, the total Altman presented in October 2025. As a deployment, it is a set of sites at different stages, announced, scheduled or operating, whose capacity figures overlap and cannot simply be added. At year-end 2025, Abilene was the only site where frontier-training racks were operating. On August 17, 2026, OpenAI agreed to secure roughly 8 gigawatts of IT capacity at a single Ohio campus, phased from 2028 with an 800-megawatt first tranche and later phases conditional on power infrastructure, permits, environmental review and financing. The April 2026 Microsoft renegotiation retired the AGI clause that had come to define a partnership dating to 2019. The commitments range over delivery dates from 2026 through 2032, financed in part by warrants, working-capital loans, and equity-coupled cloud spend. Each of the four steps in the wager can be checked against the record since January 2025. The science has, on the public record, broadly held. The cost curve has broadly fallen, with the complication that an open-weight follower captures much of the fall and that per-task frontier costs are rising. Demand has kept growing, unevenly; ChatGPT downloads slowed through April 2026, and weekly users passed one billion in August, while enterprise spending on generative AI, the market the company is narrowing toward, more than tripled in 2025 by Menlo Ventures' estimate. The conclusion Altman draws from the four steps is that the compute is worth buying, and that OpenAI is therefore worth owning. The case turns on the part that is hardest to assume. The science can be evaluated against arXiv submissions; the chase for compute can only be evaluated against deployments and against partnerships that hold or move. By 2028, on Altman's own published cadence, either revenue compounding through ChatGPT, the developer interface, the agent business and whatever follows pays for the announced infrastructure, or the partner balance sheets willing to hold it absorb the cost and the write-downs begin. Since January 2025, the capital Altman assembled in the Roosevelt Room has largely held, on terms that keep moving. Whether revenue closes the loop by 2028 is the bet.